← Back

Autonomous Equity Engine (quant_v4)

Built

Signature capability

The gate is real: the deployed sleeve failed its own pre-registered acceptance battery and was retired, with the due-diligence record written up making no forward performance claim.

What it is

The platform runs a long-only equity sleeve against a real brokerage account, with entry, rebalance and trailing-stop protection passes, fractional-Kelly sizing and a daily-loss circuit breaker. Its defining feature is the acceptance gate: thresholds are pre-registered, and combined_sleeve_v1 passed validation, went live, then failed that same battery and was retired. Entry-engine live-arming remains a deliberate manual step rather than an automatic one.

Highlights

  • Acceptance thresholds fixed in writing before any result existed: residual alpha t >= 1.96 after market and momentum factors, deflated Sharpe >= 0.95, positive net Sharpe. A missing metric fails closed.
  • Statistical circuit breaker derives its loss threshold from the win rate using a binomial-probability model, replacing arbitrary fixed-loss rules.
  • Concurrency-safe shared state: all writes route through one atomic adapter using fcntl LOCK_EX plus tempfile and os.replace, with shared locks for readers.
  • Research and live execution paths are asserted to produce identical weights to within 1e-12, so the strategy tested is the strategy that trades.
  • Roughly 20 candidate mechanisms were run through the stress gate and none survived it, which is the finding rather than a failure to report.

Tech · Python 3.10, pandas / numpy, robin_stocks (Robinhood API), fcntl file locking with atomic tempfile+os.replace writes, SQLite (WAL), systemd / launchd / cron schedulers, n8n webhook alerting, Tailscale/LAN cross-host sync, pytest.

Download brief (PDF)